Streamlining AI Agent Development with mkinf

GM GM. Welcome to what's going to be another great episode of DevNTell. So if you aren't familiar, DevNTell is a 30-minute podcast held every Friday, allowing founders, hackers, and anyone in between to come on the show and showcase their product. So today, I'm really excited to welcome Veronica, who's the co-founder and CEO of mkinf, standing for Make Inference. So mkinf is a library of hosted agents to streamline development and deployment of your agentic systems. So if you stick around for today's episode, you'll see Veronica tell us her story behind how she co-founded the company, give us an overview of the platform, and how you can get started using it today. All right, let's get into it. But first, a word from our sponsor.

Today's episode of DevNTell is brought to you by our friends at Arbitrum. Arbitrum is a technology suite designed to scale Ethereum. You can use Arbitrum chains to do all the things you do on Ethereum, but your transactions will be cheaper and faster. With over 3.8 million ETH saved, 1.69 million active wallets, and over 1.7 billion total transactions, why would you want to settle for any less? Get started today by exploring Arbitrum's comprehensive docs and tutorials linked below to help you get started launching your next dApp today. Now, onto today's show.

Hello, hello. Welcome to the show, Veronica. Pleasure to have you.

Hello, Narb. Thank you so much for inviting me on the podcast. I'm super excited for this.

Yeah, yeah. My pleasure. And, yeah, really excited for today's topic. We both know AI is all the rage, and anything that helps us develop our agentic systems easier is a godsend, so to speak. But before we kind of get into all the great stuff you have in store for us, perhaps would you like to give a brief introduction about yourself?

Yeah, of course. Veronica, from Italy, currently in the UK. I co-founded mkinf, so kind of like a vertical cloud for AI agent development and deployment. I'm the CEO, but have a background in a more technical space as well, so data science, have worked in startups as a data scientist before, and then, I mean, really passionate about startups, so decided to build my own.

Awesome. Yeah, it's a really exciting world to live in. And I guess, how did you first get interested in technology? Like, what led you ultimately to land in this space and go into AI in particular as a data scientist?

Right. So I think I was always like a nerd at core. I always liked, probably started from like a different space. I was like really good at math or like statistics, I really liked doing that kind of stuff. And but somehow I got into the wrong path at uni, started studying finance and just like did not like it at all. So I was like, okay, I need to pivot this. I need to find a way into getting somewhere out. I mean, you still study mathematics and statistics, but I realized I really wanted to learn coding. So started to learn coding and ended up with a dissertation in machine learning, but also I was like, okay, how do I actually make myself appealing to the coding entrepreneur out there, not really an entrepreneur, but like employers out there, you know, to hire me in this space. And then I was like really passionate about Formula One. So I wrote my first model, a machine learning model to predict Formula One winners. And in that season, it did like kind of better than Sky Bet or like other betting platforms. So the article I wrote about it went like kind of semi-viral. And someone in the UK read it and got an interview with them and started working as a data scientist for their startup Fintech company. So this is kind of how I got into the space.

Dang. That's amazing. And yeah, I mean, it just again showcases the power of sharing your work. Like, you never know who's watching, who's reading. So it always leads and opens up opportunity.

Absolutely, absolutely. I'd say it was a lot of luck to be fair. But no, definitely. I think like starting, I mean, at the core of AI obviously is like STEM, science and subject. So I mean, if you start from like any space, you know, like physics or mathematics or statistics, it's very likely for you to like end up there. Or even if you don't, I think it's for everyone to pick up.

100%. I totally agree. And I suppose, as you were working in the startup in the Fintech space, was there any particular moment where you're like, oh, I kind of had enough of this, I want to start my own startup?

Yeah, no, absolutely. I don't think there was like a pivotal moment. I liked working in startups I was working at. It was more about I want to have something, something of my own. And so like, you know, once I kind of realized that and putting it together, I wanted to do something on my own and the millions idea I kind of had every day. I think it was more about like handpicking the one I wanted to actually develop because I had way too many. And but then I met my now co-founder, co-founder Jonathan. He also is a hardcore nerd, one of the very first, you know, like crypto investors and miners. And fun fact, like he had like a lot of GPUs that he was like mining on like Ethereum, but then he couldn't use, like it wasn't profitable anymore to mine crypto. So we kind of just used them for our first MVP, actually. So this is where it sparked.

Right on, right on. Yeah, yeah. I mean, got to put those GPUs to use, especially now that they're so hard to get.

And I guess what kind of inspired you to land on what mkinf is today? Like, what's the reason you went down the agentic development path?

Yeah, you said something that's kind of spot on, like GPUs were really hard to find, especially like last year. And this is kind of how we started. So we saw that it was really hard to find GPUs on like AWS or GCP. And and but there were like so many, I mean, so much money being poured into the data center space. So we were like, where's all this money going if you can't find the GPUs? And we actually found out that a lot of like small independent data centers that really kind of no one knows about and do not have all the infrastructure, sorry, the software stack on top, were really running at like maybe 60, 70% capacity. And that's bad, obviously, from like a cost perspective, but also like ESG metrics. And even within the allocated capacity, maybe only 20% was being utilized. So we were like, okay, let's start optimizing this infrastructure. And so we built like a pool of GPUs like kind of going standardizing different accesses and protocols across small independent data centers and make this like kind of Airbnb of GPUs. So when you have them available, you can make them available for other people to use on demand. But the end goal, the idea was always to like simplify AI workloads. So initially we started like last year with I mean, obviously models have been generative models have been around for some time. And and it was the time to really make inference of these models and because people were like utilizing them a lot. But when you when you make inference of a model, you never really know what to expect from the usage of of the users. So you might have a peak of like thousands of users and then like down to the tens. And so having an infrastructure that is really capable of following along with that is like kind of difficult to set up. So and also from a software engineer having to deal with all those infrastructure decisions is probably something that you're not I mean, I'm not particularly passionate of like handling. So we're like, okay, let's create a platform that is able to abstract all those infrastructure decisions and make inference of AI models just really seamless, one click. But when we started, unfortunately a lot of I mean, unfortunately, fortunately, a lot of other companies and startups I mean, gained momentum like really, really quickly, got access to funds very, very fast. And and then we were like, okay, does I don't I don't think there's too many people doing this. I think there's still a lot of space. But we were like, okay, let's try to differentiate ourselves in in some way and kind of be back on top of this wave rather than following them. And we're like, okay, what's what's what do we have to make inference of? What's the next thing in AI that we can like streamline the deployment of? And there were like, well, AI agents were kind of really gaining momentum. And they were like, okay, and I don't even think the AI agent boom has like really even boomed yet. Obviously models have been around for some time, agents much less. So we're like, I think we've got like some more some more runway to prepare for like people using massively AI agents. So we're like, okay, let's reposition ourselves as a platform to kind of vertical cloud rather than focusing on AI models, focusing on one-click deploy of AI agents. But because again we're still in the phase of like building these AI agents, we do not want we do not want to give you just the one easy click deploy, we want to also help you build these AI agents. And and because AI agents, the way they're built obviously they sit on top of LLMs and then they are equipped with different tools, functions that developers often need to write from scratch or, you know, like take from GitHub repositories, but they still need to sift through like hundreds line of code before finding what they need and figure out how to integrate it. So we're like, okay, everything in coding eventually like consolidates into libraries and like more modular pieces. So let's turn this like monolithic codebase into like more modular pieces and package up those GitHub repositories into like one-line integration and host them. So like if you're looking for example of, you know, you're you're writing your agentic system and you need a function or you need to write something that is able to extract data from a website, it doesn't make sense for you now to build one from scratch. There are literally companies that are built in vertical scrapers that only do that and you're never going to have something that optimizes like someone that does that every day. So what you can do, you can go on on the mkinf library, search for use cases, so like a scraper in this case. You can obviously look at the code and modify if needed, which I think for developers is like super important having something open source that you can still inspect all the code, but you don't need to worry about, you know, okay, what kind of piece do I need to import and and copy and integrate and then host. Then all that code is like packaged up into like an SDK, SSE, API, etc. and then you can like import it into into your codebase, so like for example from mkinf import the tools that you need and they're already hosted. So when you run your code, you don't really need to take care of of deployment as every agent and tool is again hosted as a sandbox will spin up, so it's like dedicated and secure environment for each agent and user. So we go from like giving you those like building blocks, kind of like LEGO pieces, so you can build your your AI agent faster and in a more modular way, kind of like plug and play. Okay, maybe that tool is not the right one for me, I don't need to restart everything from scratch and throw everything away, I can just like pick a different tool. So this is kind of the idea and then obviously, you know, once you built it, we we want to to support you with the deployment as well and don't have to, you know, worry about those infrastructure decisions. And I think it goes I think it's kind of I'm not going to say it's like trendy at the moment, but like these like very vertical clouds like Vercel for apps or you know, Together AI again for models, like having these like vertical clouds that allow you to deploy that one thing and it's like you don't need to worry about anything else. I think it's it's kind of what we are seeing right now in the space.

Yeah, and like myself I I found myself sticking with Vercel for a lot of my side projects just because it's so easy, right? It's like Vercel isn't a sponsor of the show, not plugging them, but yeah, just mentioning that one-click magic, right, of like, oh, I built the thing, bang, now it's deployed and hosted and available to the world. It's like it's a very powerful construct.

Yeah, exactly. And and we kind of like think of Vercel for AI agents, so that's kind of the direction towards where we're going.

Beautiful. I love it. Describe describe the company in five or less words. There you go.

And so so just a review, it seems like the core components of mkinf are this cluster of GPUs that you have to host the the models that power the inference. And then you have this library of AI agents, either crafted by your by your company or people in the open source community. And then they come together to make this marketplace essentially of AI agents that people can plug and play into their agentic workflows.

Yeah, precisely, precisely. Exactly. You explained it maybe better than I did.

Well, I mean the I've got to do a TL;DR eventually, right? So but yeah, I mean it sounds like a very interesting kind of pivot from a lot of people who are still doing the the inference side, or sorry, the infrastructure side of all this hosting these models and whatnot.

I guess this is a good segue if you want to start sharing your screen to kind of walk through the the platform itself. But this this marketplace, these these agents, are they are they fully open source? Can people also upload closed source versions of their...

Yeah, this is this is actually a good question. So right now because we don't want to like we just launched the product like two months ago, so we don't really want to overwhelm the users and the platform with like a million tools and agents. So it's still like kind of a curated library by us, like covering main use cases like browsers, scrapers, or integration with different platforms like Slack, Notion, or Superbase, for example. And these are all open source. As we grow obviously, we I mean, we've already started taking contributions from from the community that want to host their their agents. And they can either do that obviously privately or publicly in the in the library. And then they can kind of decide whether they want to like obviously, you know, show the show the code or or not. And and in that case obviously, you know, by you by by our developers in the community maintaining maintaining the agent and maintaining the tools, we we have we have in place, little spoiler, we have in place obviously a remuneration like program obviously if you if you maintain the agent and everything you get a chunk of of the hosting and everything.

Yeah, I was going to ask if there's any monetization opportunities. So it seems like as long as your agent is maintained, kept up to date, then you get a piece of the pie.

Yeah, of course, of course. No, and I think it's obviously common saying there's no money in the in the open source, but obviously by putting together the hosting and then everyone contributing, we want to like kind of give them a piece of the pie that we are we are building.

Awesome. Love to hear it. And just curious, your co-founder being or having a background in crypto, have you decided at all if there's going to be any sort of like token around the the platform? Maybe this is a roadmap thing or alpha thing. You don't have to say, but...

It's to be to be. We'll see. We'll have to discuss it for the roadmap. There's a lot of obviously like a lot of stuff to do. And obviously like people I feel like understanding that this is a cool space to be in. And obviously with for example, MCP registries out there, MCP being Model Context Protocol, a new protocol for building AI agents and and right now there are like a lot of servers out there that are like being kind of collected as you were saying like registries or or marketplaces, but you still need to download these servers and host them. And obviously the next natural step for these registries would be to host these these servers. And I mean, we're already there, but to really stay ahead of the curve you need to you need to integrate emerging technologies before they even emerge. That's kind of that's kind of the thing. I feel like developers like shiny things and and they're the first to pick up on on new trends. And obviously when building for developers, you really need to have that kind of mentality.

Yeah, I 100% agree. Well said. And yeah, I I'm really curious to kind of see what the platform's all about. So...

Yeah, of course. I'll share my screen. Let me present. And for those of you watching on the live stream, if you have any questions for Veronica, please post them and we will get those answered for you.

All right. I hope you can I hope you can see my screen. As I mentioned, we launched like two months ago, so this is going to be looking way nicer soon in the future. As I mentioned, I mean, once you once you sign up, you're you're going to be prompted to create your profile and and create your API key to actually utilize our platform. As I was saying before, we've made we've integrated and hosted different agents and and tools ranging from I mean, browser scrapers or various platform integrations. The way each repository is structured, I'll just show you one, is like kind of very similar to to GitHub. I mean, building for developers, we kind of want to make the want to build something for them that they're already familiar with. You can inspect the code and and the files in here and you also have a link to the GitHub page here. And you got the Readme here and obviously the the main let's say there are contributions despite I don't want to minimize it, but it's just kind of this like purple box that package up the code, either that being a GitHub repository or an MCP. And you can, I'm going to probably zoom in a little bit, and you can like integrate it with a Python SDK, SSE, or HTTP request. You also got I mean, for this particular agent you only got one action that it can perform, but for others you can see the actions that it can perform in more detail with the input and the output. Now, once you've selected the the tool that you want to be utilizing, this one for example is able to extract like schemas and and content from from GitHub. So I'm going to be using this.

And I'm going to be copying this into into my codebase. This is like a super simple graph. I hope you can still see my screen. This is like a super simple graph with an OpenAI model underneath here and a very, very basic logic, so a few edge nodes here and there. And you can so what I'm doing here, I'm like importing from our library the tools that I'm needing, so in this particular case only this GitIngest one. I can potentially list multiple ones here. Now, once I wrote the the graph and the logic that I want here, we've just built a super basic Streamlit UI on top so that I can just like see the tools that I'm I'm utilizing and I can just chat with it. So for example, I can ask it get me the readme of this repo mkinf mkinf io github. And obviously depending on well, first my internet connection and the the task that you've asked it, it's going to take like different time, different amount of time. What takes when I when I when I run the code, it took like less than a second to spin up a sandbox. Oh, actually here it gave me the response. Spin up a sandbox. So this is not running locally on my machine, it's already like running in the cloud. So it spun up a sandbox, it's a dedicated environment for each tool. So if I had multiple tools, it would be like multiple sandboxes and just dedicated for for me.

Now, we also made an integration with Cursor for all the vibe coders out there. I'm going to show you I'm going to show you so yeah, so basically what you do, you just actually there's a new way now to import a new MCP server on on Cursor, but the previous way you can just like list multiple in a JSON file the SSE that you can that you can find here. Now I have a few available ones, so I'm going to spin them up. I'm actually probably going to be using only a couple of these. Anyway, and I need to wait for it to spin up and I can just chat with it. So for example, I can ask it use playwright, which is actually a new tool, playwright, to go to mkinf.io and for example get all the blog post, I don't know, in a JSON format. Hopefully the demo gods are with me. Now here it's understanding, I mean, I actually already prompted it to use a playwright, but it's calling the MCP tool and it's understanding which tool here on playwright, you got 25 available tools, so it's understanding which sorry, not tools, tools, actually. Yes, tools, actually. Which action to to perform. It's navigating in the website, understanding what it needs to extract, go on the on the section of the blog post, getting a browser snapshot. And now obviously depending change a couple of things.

I've identified the section thoughts on AI. So here did I not say JSON format? Okay, I don't know. I would like to include a JSON. Okay, I would like can would you like to the JSON? I would like the JSON to include title, subtitle and date, for example. Anyway, this one went way better before when I did it. Sorry about that, but yes, now it's giving a JSON with the article, the most recent article that we wrote, definitely should update more since date January. But yeah, so this is this is what I what I got. So it was able to navigate to the website, get the content that I wanted and and give me and give me a response. This is again mainly using Cursor and so you're just able to chat with it. Obviously the real power comes from like putting more of these together and making them communicate with each with each other and creating like a again a cool graph system for integrating multiple multiple systems and and endpoints.

Now, maybe a cool one I wanted to show you so that you can actually see how to get started with mkinf. I can just say use playwright, I clearly have a difficulty writing playwright. Use playwright to go to mkinf.io. Then I say in the navbar click on the docs and extract the first step to get started. Okay, let's see what it gets me. Calling MCP tool browser navigate. Again, this is all hosted as well. Like the the real power is that you already it's production ready tools and agents that we are that we are making available for people to use. Now it's utilizing a few, I think it went into a loop that. Okay, the first step to get started, create a mkinf account. You go on the mkinf so hub.mkinf.io, sign up and obviously during the beta the beta period all accounts receive unlimited free credits. So yeah, this is this is how to get started. I hope I'm still there.

Yeah, and that's that's great because that was one of the questions I was going to ask. So you got you got AI to do it for you. There you go. Glad it worked. No, I'm kidding. Yeah, no, you'd be surprised how many people have issues with demos. So yours yours went great, so yeah.

So firstly, that's really, really cool. And I wanted to highlight, like, it took about I think it was three or four lines of code for you to import the library, pull in the model, and then basically plug it into the rest of your existing workflow. And I think that's a really powerful construct just because there's like so many different AI frameworks out there and you guys are doing a very a great job in just making it really stupid simple for people to plug and play with any model that's in the marketplace there.

Yeah, absolutely. And and one I want to be absolutely clear about is that I mean, you mentioned so many frameworks out there like LangChain, CrewAI, SmaleAgents, etc. and we want to be obviously to start with as open as possible. Like, you can utilize whatever framework you want onto our platform, whether that being again LangChain, CrewAI, etc. so yeah.

And can you speak a little bit about how someone would basically bring in their AI agent into the marketplace? Is there a import process or if they or do they just fork a starter template?

Yeah, yeah, of course, of course. So right now onto onto our platform you can there's a section for like new repository and you can add it there. We're working on a UI that's going to be like ready available very soon, but we are also like super active on our, you know, Slack and Discord communities. So if you have any, like, you can just chat with us and I mean, if there's even if there's no UI yet for like putting stuff into production, you can still put stuff into production through us and we'll still update you on the monitoring of at the utilization and everything, but the UI is coming really soon.

Yeah, excellent. Great to hear. And obviously you're very gung-ho on AI agents, as am I, and we basically will see them evolve over the next few years. And just want to get your perspective of how you see these AI agents evolving and how they're going to be part of people's day-to-day whether they realize it or not. Just curious on your thoughts.

Yeah, no, absolutely. I think the this like tool integration is definitely going to be like the norm. They'll take actions, run code, query APIs and in terms of I think like there's like difference to draw here between like AI agents and models. Obviously with like models they're like super generalist, they need lots of resources to be trained. So I feel they're still going to be like in the hands of like companies that have those resources to like train massively and make really good models. But because all the data in the world doesn't fit into a single LLM and and also I mean, that data is like proprietary, they do not have access to it, what AI agents will enable is like this this data like kind of like empowered with AI. So what I'm imagining, I mean, for certain AI agents like browser scrapers obviously probably be condensed into like a few powerful and and really good players. But when it comes to agents that are like let's say experts or like have access to specific data into like the medical field, the law, finance, etc. and are owned by big enterprises, what what I'm expecting is like domain-specific agents, so like a verticalization of all these agents and will be where we'll get access to this data. I mean, we're probably not going to see the data but still will get access and we'll be able to chat with them. So this is kind of where I see the the AI agent market going. And I don't and I don't think absolutely we are there yet because like these enterprises right now they're just like incorporating LLMs right now and they just they just trying to they just understanding, you know, how powerful these can be and and maybe maybe in a few I'm hoping months, but I don't know if it's if it's going to be months they'll they'll start using AI agents as well.

Yeah, I I would be surprised if they didn't. I mean, like if you can spin up an AI agent to do the mundane tasks that would take up like an hour or two of your time, like why not? You can use that hour or two to do more productive work. But yeah, I mean I share the same viewpoint and I think yeah, AI agents are going to be a staple whether people realize it or not.

And just curious, is there anything you can share around the mkinf's roadmap? I know you said UI improvements and whatnot, but anything further you can share as well?

Yeah, yeah, of course. I mean, we're still small team, so we definitely want to get more more power on the team to to build faster. Definitely one thing that I can say it's definitely on the roadmap is the use cases that we're going to offer on to our library, like expanding a lot more, especially with there are like certain use cases that are more difficult to integrate, like with memory tools, for example, like we want to tackle those. Or also offer I mean, right now you kind of just saw text input and output, but having different media as well, so for inputs and outputs, voice, video, image, etc. so that's that's definitely that. With MCP despite being kind of like being validated as a as a protocol, I think I still think there's a lot of work to do around that, especially around authentication and everything. So we'll see how that evolves and we'll obviously like keep up with with that. But then also because the end goal is is the is the is the one click deployment really really being able for you to not think about what's running underneath. That's really the end goal.

Amazing. And yeah, I'm looking forward to seeing mkinf continue to evolve. It already looks really really great. If you're a developer looking to tinker with AI agents, give mkinf a try. The link to their website is in the YouTube description as well as the Spotify description if you're watching this after the fact. And with that, I just want to thank you so much, Veronica, for taking the time out of your day to come chat with us.

Thank you, Narb. It was a pleasure to be here and show you guys what we're building. Get in contact if you like it or like any feedback, happy to chat with anyone.

Yeah, yeah. I was going actually ask you, so what's the best way for people to reach out if they want to get in contact?

Any, really. Like I'm on LinkedIn. I am if you want I can give you like our Slack channel and you can join us on on our platform. We post like demos on there, so like happy to to get in contact with anyone.

Amazing. Yeah, we will share that share those links in the description. And yeah, just with that, I want to wish everybody a very happy Friday, happy weekend wherever you may be and we'll catch you back here for another great episode of DevNTell next week. All right, everyone. Cheers. Thank you. Cheers.